Linear regression

Linear Regression

When to use this skill

Use for continuous outcome prediction or explanation where interpretability matters. Triggers:

• "Predict revenue / session time / NPS"
• "What drives <continuous metric>?"
• "Linear regression for…"
• "OLS"
• "Explain the variation in…"

For binary outcomes use logistic-regression. For time-to-event use survival-analysis. For pure prediction with non-linear effects, fit linear first as baseline, then suggest gradient boosting.

Required inputs

Input	Why it matters
Continuous target	What you're predicting (numeric, ideally not heavily skewed)
Feature set	Predictors
Observation grain	Per user / per session / per geography
Train/test strategy	Temporal split for production, random for exploratory
Goal	Pure prediction, coefficient interpretation, or both?

Workflow

1. Audit the data (data-quality-audit skill).

2. Inspect the target distribution.

   import matplotlib.pyplot as plt
   df[target].hist(bins=50)
   df[target].describe()
   

   • Symmetric, finite variance → OK for OLS
   • Heavily right-skewed (revenue, time-on-page) → consider log(1 + target) transformation
   • Heavy-tailed with extreme outliers → robust regression or winsorize

3. Check for outliers at p99 and p99.9. Decide:

   • Cap at p99 (winsorize)
   • Drop with documented justification
   • Keep and use robust regression

4. Verify no leakage (same checks as logistic-regression).

5. Split temporally (or random for exploratory).

6. Fit baseline OLS with all features. Use statsmodels for coefficient inference, sklearn for production fitting.

7. Check assumptions (diagnostics matter more than for logistic):

   • Linearity: residuals vs predicted should look random
   • Homoscedasticity: residual variance constant across predicted values
   • Normality of residuals: Q-Q plot
   • No multicollinearity: VIF < 5 (warn at 5-10, fail at > 10)
   • Independence: no autocorrelation (for time-series)

8. Evaluate on holdout:

   • RMSE (in original units — interpretable to stakeholders)
   • MAE (more robust to outliers)
   • R² (variance explained — but easily inflated, don't worship)
   • Out-of-time R² is the real test

9. Interpret coefficients.

   • Standardized coefficients to compare features
   • Report 95% CIs (statsmodels gives these)
   • Watch for sign flips when adding/removing correlated features (multicollinearity sign)

10. Write the readout.

Output format

# Linear Regression: <outcome>

## Setup
- Outcome: <target> (unit: <e.g., dollars per user>)
- Distribution: mean <X>, median <Y>, p99 <Z> (transformation: <none | log(1+y) | winsorize@p99>)
- Observation unit: <user / session / geo>
- Sample: <N_train> train, <N_test> test (split: <temporal at YYYY-MM-DD | random>)
- Features: <N> features

## Performance (test set)
| Metric | Value | Notes |
|---|---|---|
| RMSE | 18.4 | Target std = 27.1 |
| MAE | 12.1 | |
| R² | 0.47 | |
| R² out-of-time | 0.42 | |

## Coefficients (top by |effect|, standardized)
| Feature | β | SE | t | 95% CI | β std | Interpretation |
|---|---|---|---|---|---|---|
| tenure_months | +1.84 | 0.12 | 15.3 | [+1.60, +2.08] | +0.41 | +1 month → +$1.84 revenue |
| has_team_plan | +24.10 | 1.42 | 16.9 | [+21.3, +26.9] | +0.38 | team plan → +$24.10 revenue |
| device_mobile_pct | -0.32 | 0.04 | -8.0 | [-0.40, -0.24] | -0.18 | 1pp more mobile → -$0.32 revenue |
| ... | | | | | | |

## Diagnostics
- Residuals vs predicted: <looks random | shows funnel pattern (heteroscedasticity) | shows curvature (non-linearity)>
- Q-Q plot: <residuals roughly normal | heavy tails>
- VIF: <max VIF = 3.1 (OK) | max VIF = 12.4 (multicollinearity warning)>
- Autocorrelation: <Durbin-Watson = 1.94 (OK)>

## Caveats
- <e.g., predictions outside the training range of feature X are extrapolation; not reliable>
- <e.g., model assumes effects are linear; non-linear effects detected for feature Y>
- <e.g., R² inflated by 0.07 vs out-of-time R²; mild overfitting>

## Next steps
- <e.g., target seasonal effects with interaction terms>
- <e.g., investigate residual cluster around segment Z>
- <e.g., if pure prediction matters, benchmark vs gradient boosting>

Validation checks

• Target distribution inspected; transformation decision documented
• Outliers checked and policy stated
• No leakage in features
• Residual diagnostics ran (linearity, homoscedasticity, normality)
• VIF checked for multicollinearity
• CIs reported alongside coefficients
• Out-of-time R² compared to in-sample R²

Edge cases & failure modes

• Heavy-tailed target (revenue): OLS gives undue weight to outliers. Use log-transform or quantile regression.
• Zero-inflated target: many zeros + continuous positives (e.g., revenue with many free users). Use two-part model: P(zero) via logistic, then OLS on positive values.
• Multicollinearity: coefficients become unstable. Drop one of correlated pair, or use ridge regression.
• Non-linear relationships: residuals show U-shape vs predicted. Add polynomial terms or use a tree-based model.
• Heteroscedasticity: residuals fan out at higher predicted values. Coefficient SEs are wrong. Use HC3 robust SEs (cov_type='HC3' in statsmodels).
• Time-series autocorrelation: residuals from consecutive periods correlated. SEs are too small. Use Newey-West SEs.

Scripts

• scripts/fit_linear.py — End-to-end OLS fit + diagnostics + report.

python scripts/fit_linear.py \
    --input data.csv \
    --target revenue_per_user \
    --features feature_list.txt \
    --transform log1p \
    --split-by month --split-date 2026-03-01

Related skills

• data-quality-audit — run before fitting
• logistic-regression — binary outcomes
• causal-inference — when you need effect estimation, not just association
• stakeholder-readout — for packaging the model output